Non-rigid Registration for Large Sets of Microscopic Images on Graphics Processors

Abstract

Microscopic imaging is an important tool for characterizing tissue morphology and pathology. 3D reconstruction and visualization of large sample tissue structure requires registration of large sets of high-resolution images. However, the scale of this problem presents a challenge for automatic registration methods. In this paper we present a novel method for efficient automatic registration using graphics processing units (GPUs) and parallel programming. Comparing a C++ CPU implementation with Compute Unified Device Architecture (CUDA) libraries and pthreads running on GPU we achieve a speed-up factor of up to 4.11× with a single GPU and 6.68× with a GPU pair. We present execution times for a benchmark composed of two sets of large-scale images: mouse placenta (16K × 16K pixels) and breast cancer tumors (23K × 62K pixels). It takes more than 12 hours for the genetic case in C++ to register a typical sample composed of 500 consecutive slides, which was reduced to less than 2 hours using two GPUs, in addition to a very promising scalability for extending those gains easily on a large number of GPUs in a distributed system.

Parallel Automatic Registration of Large Scale Microscopic Images on Multiprocessor CPUs and GPUs

“…The process of registration is the focus of our high performance computing effort in this paper, which extends our previous work on a singleprocessor platform [2] to make use of massive parallelism. For a mouse mammary sample composed of 500 slides, it took more than 181 hours for our C++ code to accomplish the registration process on a high-end CPU.…”

Real-Time Non-rigid Registration of Medical Images on a Cooperative Parallel Architecture

“…Levin et al [7] implemented a high-performance Thin Plate Spline (TPS) volume warping algorithm that accelerated the application of the TPS nonlinear transformation by combining hardware-accelerated 3D textures, vertex shaders, and trilinear interpolation. Antonio et al [8] used polynomial mapping as non-rigid transformation and achieved a factor of 4.11 speedup with a single GPU and 6.68 with a GPU pair over CPU-based NRR. Vetter et al [9] implemented non-rigid registration on a GPU using mutual information and the Kullback-Leibler divergence and reported GPU performed up to 5 times faster per iteration than the CPU implementation.…”

“…Recently, some groups implemented it on Graphics Processing Units (GPUs) [7], [8], [9], [10], [11]. However, up to now there were no reports on accelerating NRR using the cooperative architecture: multicores and GPU, which is widely available in commodity PCs.…”

Power-efficient medical image processing using PUMA

“…For instance, in [11], the authors port "large-scale, biomedical image analysis" applications to multi-core CPUs and GPUs, and compare different implementation strategies with each other. In [21], the authors study image registration and segmentation and accelerate those applications by using CUDA on a GPU. In [24], the authors use both the hardware parallelism and the special function units available on an NVIDIA GPU to dramatically improve the performance of an advanced MRI reconstruction algorithm.…”